Useful blunders: Can automated speech recognition errors improve downstream dementia classification?

Area of Science:

• Computational linguistics
• Neuroscience
• Artificial intelligence in healthcare

Background:

• Alzheimer's disease (AD) diagnosis relies on cognitive assessments, often involving language analysis.
• Automatic Speech Recognition (ASR) systems are increasingly used for transcribing spoken language.

Purpose of the Study:

• To investigate the impact of ASR errors on dementia classification accuracy.
• To determine if imperfect ASR transcripts can aid in distinguishing Alzheimer's disease from healthy cognition.
• To evaluate the effectiveness of ASR-generated transcripts in the "Cookie Theft" picture description task.

Main Methods:

• Experiments utilized various ASR models with post-editing for transcript refinement.
• Both imperfect ASR and manual transcripts served as input for dementia classification models.
• Comprehensive error analysis was performed to compare model performance.

Main Results:

• Imperfect ASR transcripts outperformed manual transcriptions in classifying Alzheimer's disease.
• ASR-based models surpassed previous state-of-the-art performance.
• ASR errors were found to contain valuable dementia-related linguistic cues.

Conclusions:

• Imperfect ASR transcripts capture linguistic anomalies associated with dementia, enhancing classification accuracy.
• The synergy between ASR and classification models highlights ASR's potential in cognitive impairment assessment.
• ASR shows promise as a tool for clinical applications in evaluating cognitive health.